Hi Dat, thanks for the quick response. For sure retraining would be the optimal solution, but I was looking for something faster just to see the results, I really think that can be interesting for anybody to control whatever they want to detect. I finally came up with a dirty/quick solution like this:
_, not_person_elements = np.where(classes != 1)
for i in not_person_elements:
scores_np[i] = 0.01
The where function returns the a size two tuple, where the first element is an array of 100 zeros and the second (not_person_elements) are the positions of the detected classes that are not “Person” (that’s the reason of the  in the loop). After that I give the corresponding element in the score array a small value in order to not display it. Not elegant but seems to work.
Thanks again and keep going with your great articles, I’ll look for them for sure.